Lesson 3 of 6 · Agentic Processes
Agent mode and browsing
Research reads the web. Agent and browsing modes act on it: open sites, fill forms, compare carts, click buy. Here is what that means, where each app stands as of August 2026, and the moments a good agent must stop and ask you.
Lesson 2 covered features that read. This lesson covers features that do. A browsing agent is the same plan, act, check loop from lesson 1, but the actions are clicks and keystrokes on real websites instead of searches. It can log into your supplier's portal, find last month's order, and put it in the cart. It can open six rental sites, set the same dates on each, and build you a comparison. It can fill out a form from a PDF you give it. It is the most useful and the most dangerous thing in this module, and the product names for it changed as recently as this month.
What "it clicks for you" really means
Under the hood there are two ways to build this, and it helps to know which one you are using because they fail differently.
- In your browser. A browser extension lets the AI see the page you are looking at and click inside it, using your existing logins. You can watch it work in real time, and you can grab the mouse the moment something looks off. Claude in Chrome and Gemini's auto browse in Chrome work this way.
- In a cloud browser. The AI gets its own browser on the company's servers. It is not logged into anything of yours unless you sign in for it, and it keeps working after you close the tab. You see a video-like feed or a log rather than your own screen. OpenAI's browser agent work is moving in this direction as of August 2026.
Either way, each step is the loop: the model looks at the page (as text, a screenshot, or both), decides on one action, the harness performs it, the model looks again. A twenty-step task is twenty rounds of that. It is slower than you clicking, and much faster than you clicking on six sites in parallel.
Step through an agent booking a rental. Should it stop and ask?
Find me a side-by-side UTV rental in Hurricane for this Saturday, four seats, under $400 for the day, and book it.
Step 1 of 6
Searches for UTV rentals in Hurricane and opens the top four rental sites in separate tabs.
Illustrative. Business names, prices, and site layouts are invented. The pattern of where an agent should pause is the real lesson.
The three apps, as of August 2026
This section will age faster than anything else on this site. Names first, then the pattern that outlasts them.
- ChatGPT. The feature called ChatGPT agent (often called agent mode) was retired in early August 2026. Its abilities are now split: ChatGPT Work, launched July 9, 2026 on Plus and above (not Free or Go), handles long multi-step jobs that end in a document, spreadsheet, or presentation, with a plan you approve before it starts and check-ins along the way. Browser-driving work is moving to a cloud browser inside ChatGPT and the desktop app. OpenAI's standalone Atlas browser shut down on August 9, 2026. If you had a workflow that leaned on the old agent mode, test it again rather than assuming Work covers it.
- Gemini. Google's browsing agent is called auto browse, and it lives inside Chrome for Google AI Pro and AI Ultra subscribers in the US. It is built to pause and ask before a purchase or a social media post, and with your permission it can use Google Password Manager to get through sign-ins. Ultra subscribers also get Gemini Spark, an always-on agent that runs tasks on a schedule and is designed to check with you before major actions.
- Claude. Claude in Chrome is a browser extension available on all paid plans (Pro, Max, Team, Enterprise), still labeled beta. It reads, clicks, types, and fills forms using your existing logins. It asks for explicit approval before irreversible actions like purchases or sharing personal data, and runs a separate check comparing each consequential action against what you originally asked. As of mid-August 2026 the extension's side panel became a full Cowork session for Max and Team, with Pro following. It is Chrome only and not on mobile.
The pattern that survives the renames: every mainstream browsing agent is now built to stop at money, at logins it does not have, and at anything it cannot undo. If a tool you are trying does not do that, it is not ready for your business.
Logins, payments, and the moments it must stop
Think about what you would let a brand-new assistant do on their first day with your accounts. Read your supplier portal, sure. Fill a cart, fine. Type your card number from memory and click Place Order? No. Same rules for the agent. The four stop-and-ask moments:
- Sign-ins. The agent should use a login you already have open or hand the screen to you. It should never be given a password in chat. Claude's 1Password integration and Gemini's Password Manager option exist precisely so the agent fills the credential without ever seeing it.
- Money. Card entry, Place Order, Confirm Booking. Every one of these apps is designed to pause here. Let it.
- Sending. Emails, messages, social posts, form submissions. Anything that leaves your account and reaches another person. Review before it goes.
- Deleting or changing records. Cancelling an order, editing a listing, changing an account setting. Ask for a summary of the change first.
Put those in the brief. "Stop before any payment, any form submit, and any login, and show me what you are about to do." The apps have their own guardrails, but stating yours costs nothing and catches the gaps.
What it still cannot do well
A short honest list, so you are not surprised:
- Sites that fight automation. Some sites block anything that looks like a bot, throw up puzzles, or hide content until you scroll. The agent will tell you it is stuck. Finish that step yourself.
- Judgment calls in the middle. "Book the cheaper one unless the reviews are bad" is a judgment. Decide what "bad" means in the brief or expect to be asked.
- Speed. A task you could do in three minutes may take the agent five. The win is tasks you would not do at all, or that run while you do something else.
- Your phone. As of August 2026 the in-browser agents are desktop Chrome features. Cloud browser and scheduled agents are the mobile path, and both are still early.
None of this makes browsing agents a gimmick. It makes them a new hire: useful from day one on bounded tasks, trusted with more as they earn it. The rule of stakes from AI Basics lesson 9 applies to actions the same way it applies to facts.
Try this yourself
Start with a read-only task where nothing can go wrong, and where the agent stopping at the right moment is part of what you are testing. Use whichever browsing feature your plan includes.
Open the websites of three equipment rental companies in the St. George area and find the daily rate for a 3,000 pound mini excavator this Friday. For each, note the rate, any deposit or delivery fee, and the cancellation policy. Rules: do not log into anything, do not enter any personal or payment details, do not submit any form. If a site asks for any of that to show a price, skip it and tell me. Deliver: a short table, and a line telling me which site you skipped and why.
Watch it work. Note where it hesitated, where it asked, and whether it respected the rules. If it did, next time give it one step further, such as filling a quote form and stopping before submit. That is how you build trust with a person, and it is how you build it with this.
Last updated August 24, 2026